Towards reconstruction of Pulsed-wave Doppler signals from Non-invasive fetal ECG

Aman Verma; Deva Satya Sriram Chintapenta; Saikat Mujumder

REVIEW 4 major objections 5 minor 1 cited by

Towards reconstruction of Pulsed-wave Doppler signals from Non-invasive fetal ECG

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Non-invasive fetal ECG can be mapped to pulsed-wave Doppler envelopes, a feasibility study argues.

desk verdict Sincere first pass at a new task, but the paper's own r-values and own figures show it hasn't demonstrated feasibility yet. read the letter →

arxiv 2501.16651 v1 pith:MEK4AC7M submitted 2025-01-12 eess.SP physics.med-ph

Aman Verma , Deva Satya Sriram Chintapenta , Saikat Mujumder This is my paper · ORCID

classification eess.SPphysics.med-ph

keywords fetalECGpulsed-waveDopplersignalreconstructioncross-modalitylearningUNetnon-invasivemonitoring

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that non-invasive fetal ECG (NI-fECG) can be used to reconstruct pulsed-wave Doppler (PwD) signal envelopes, which are richer indicators of fetal cardiac blood flow but require expensive ultrasound equipment. The authors build a preprocessing pipeline that turns PwD images into time-series envelopes and extracts fECG from abdominal ECG, then train a residual UNet-style network (PwDRecNet) to map fECG segments to PwD envelopes. Their numerical results are weak, with the best Pearson correlation reaching only 0.0453, but the pattern across ablations—matched polarity and configuration improve reconstruction—leads them to conclude that the reconstruction is feasible with better alignment and a stronger network. The significance is that, if the feasibility holds, cheap and non-invasive ECG could approximate Doppler-derived fetal cardiac health information.

What carries the argument

The central object is the PwDRecNet architecture: a 1D convolutional encoder-decoder with three encoder blocks and three decoder blocks, where each block contains three residually connected 1D convolutions, followed by pooling in the encoder and upsampling in the decoder, with a final convolution that outputs two channels representing the upper and lower PwD envelopes. It is trained with mean-squared error, RMSprop at a learning rate of 1e-3, batch size 128, and 50 epochs. Its role is to learn a regression from fixed-length NI-fECG segments (resampled to 284 Hz) to corresponding PwD envelope segments, capturing temporal contexts at multiple scales through the UNet-style pooling and the residual connections that preserve fine-grained features.

What would settle it

A decisive test would be to take a dataset where fECG and PwD are recorded simultaneously with known temporal alignment, or align them manually using cardiac-cycle fiducial marks, then train PwDRecNet with the best configuration and measure Pearson correlation on held-out records. If the correlation remains near zero (for example, r < 0.1) even with explicit alignment and matched polarity, the paper's feasibility claim would fail; if it rises substantially (r > 0.5), the claim is strongly supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that PwD signal reconstruction from NI-fECG is feasible: a residually connected UNet-style 1D CNN can map denoised, resampled fECG segments to PwD envelope waveforms, given the right wave configuration (EA+) and envelope choice (upper envelope or PCA-compressed single channel), and a stronger network than the current prototype. The evidence is indirect: correlation values are near zero, but consistent improvements when input fECG polarity matches output PwD orientation (r rises from 0.0245 for grouped configurations to 0.0453 for upper-envelope EA+) indicate that the inputs contain learnable physiological correspondence, and that misalignment rather than absence of information is the main barrier.

Load-bearing premise

The load-bearing premise is that after independent resampling and fixed-interval windowing, each fECG segment and its paired PwD envelope segment represent the same fetal cardiac cycles, so that a mapping learned between them is meaningful. If the two signals are not actually time-aligned, the network cannot learn a genuine physiological correspondence, and the near-zero correlations reflect misalignment rather than a fundamental infeasibility.

Editorial extensions

If this is right

If alignment and configuration are fixed, a better network can reconstruct PwD envelopes from NI-fECG, potentially enabling low-cost fetal cardiac blood-flow monitoring without ultrasound.
Matching input fECG polarity to output PwD orientation is a key factor, with EA+ configurations and positively polarized fECG reconstructing best at r equal to 0.0453.
PCA-compressing the upper and lower envelopes into a single channel improves reconstruction for EA+ (r equal to 0.0452 at t=0.75 s), suggesting that reducing output dimensionality helps the model.
Simple regression models such as linear, Ridge, and Lasso fail to learn the mapping, indicating that the relationship is nonlinear and requires a deep architecture like PwDRecNet.
Creating a joint envelope from upper and lower PwD envelopes that corresponds to fECG in time and periodicity is a necessary next step for improved reconstruction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the feasibility holds, a clinically useful system would need explicit temporal alignment, for instance using fetal heart rate from fECG to gate Doppler cycles; the paper's fixed-interval windowing is likely too crude.
The near-zero correlations suggest the current preprocessing discards phase information, so incorporating a differentiable alignment layer or training with a time-shift-invariant loss could test whether the apparent infeasibility is an artifact of misalignment.
The finding that polarity matching matters hints that the network may be learning a simple sign correlation rather than a rich physiological mapping, and a sign-invariant feature extractor might generalize across fetal orientations.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Two things to know. The task is genuinely new—nobody has tried to reconstruct pulsed-wave Doppler envelopes from non-invasive fetal ECG—and the paper is honest about its preprocessing and about the mismatch between the signals. But the central conclusion, that reconstruction is feasible with the right configuration and a better network, is not supported by anything in the paper. The best Pearson r is 0.0453; most ablations are near zero or negative; Figure 4 shows nearly flat predictions and explicitly says the signals are misaligned. That is a near-flat predictor, not a learned mapping.

What the paper does well: it defines a concrete cross-modality task, uses a public dataset (NInFEA-DB), builds a reproducible-looking preprocessing chain for extracting PwD envelopes from images, and runs a broad ablation across wave configuration, envelope choice, polarity, input length, and regression baselines. It also openly lists the reasons performance is low, including the lack of explicit time alignment. That transparency is real and gives the work value as a negative pilot.

The soft spots are mostly in the interpretation. The peak r is selected post hoc across many ablation cells, with no confidence intervals and no trivial baseline (e.g., predicting the mean envelope). More importantly, Section 5.1 and Figure 4 both concede that the fECG and PwD pairs are not time-aligned. If the target envelope for a given ECG window can be from a different cardiac cycle, no supervised model can learn a phase-resolved mapping. r≈0.045 is exactly what you'd get from a flat predictor. The conclusion that feasibility is shown is therefore not just optimistic; it's contradicted by the paper's own evidence.

I'd read this as a useful negative result in need of reframing. If the authors rework the alignment (or at least test whether the pairing holds), add error bars and a mean-prediction baseline, and present it as "we tried, here's why it fails, here's what's needed," it could be a legitimate pilot study. As written, the feasibility claim is unsupported.

For peer review: I'd send it to a referee rather than desk reject, because the task framing and negative data are worth a critical look and the fixable flaws are methodological, not fatal. But I expect a major revision or rejection unless the framing changes.

Referee Report

4 major / 5 minor

Summary. The paper asks whether pulsed-wave Doppler (PwD) envelope signals can be reconstructed from non-invasive fetal ECG (NI-fECG), and proposes PwDRecNet, a residually connected UNet-style 1D CNN. The pipeline preprocesses NInFEA-DB data, extracts upper/lower PwD envelopes from Doppler images, resamples both signals to 284 Hz, splits them into fixed-length windows, and trains regression models to map fECG windows to PwD envelopes. The authors report ablation studies over input length, batch size, wave configuration, envelope selection, fECG polarity, PCA-compressed outputs, and regression baselines, using Pearson correlation r as the main metric. The best reported result is r=0.0453 (EA+ upper envelope), and Figure 4 shows predicted envelopes that are nearly flat and visibly misaligned with the ground truth. The paper nevertheless concludes that reconstruction is feasible if the 'right configuration' and a better network are used.

Significance. If a reliable mapping from NI-fECG to PwD envelopes existed, it could make fetal cardiac Doppler information available from inexpensive and convenient ECG recordings, which would be clinically significant. The paper is the first to pose this specific cross-modality reconstruction problem, and it evaluates on a public multimodal dataset (NInFEA-DB), which is a useful choice. The authors also deserve credit for openly reporting the poor alignment and near-flat outputs in Section 5.1 and Figure 4. However, the experimental evidence does not establish that reconstruction is feasible: the best correlation is effectively zero, no null baseline is reported, and the paper's own figures and text show that the learned predictions do not track the target envelopes. As a positive feasibility claim the paper is not convincing; as a documented negative result it could still be valuable, but that is not the framing used in the title, abstract, or conclusions.

major comments (4)

[Sections 5.1, 5.7, and 6] The manuscript's central conclusion that reconstruction is feasible is contradicted by its own evidence. Section 5.1 states that 'time alignment between the PwD and fECG signal is not explicitly present,' and Figure 4 shows predicted envelopes that are 'nearly flat' and fECG/PwD pairs that are 'not aligned properly.' With misaligned targets, a supervised model cannot learn a phase-resolved mapping, and a nearly flat predictor would produce exactly the reported values around r=0.0453. The paper never compares against a trivial baseline such as predicting the mean envelope, so the reported correlations cannot be distinguished from noise. The feasibility conclusion in Sections 5.7 and 6 is therefore unsupported.
[Section 3.1] The preprocessing pipeline resamples the fECG and PwD envelopes to 284 Hz and then splits them into fixed time intervals, but it does not include any synchronization, delay-estimation, or cardiac-cycle gating step. The text calls the resulting pairs 'respectively coherent signals,' yet Section 5.1 and Figure 4 directly contradict that characterization. Without a demonstration that each fECG window and its paired PwD envelope correspond to the same cardiac cycles, the supervised learning setup itself is not valid. This is a load-bearing issue for the central claim.
[Tables 1-5] The best result r=0.0453 is selected post hoc from a large grid of ablation cells (input lengths, batch sizes, wave configurations, envelope choices, polarities, PCA options), with no confidence intervals, no per-seed variance, and no multiple-comparison correction. Many entries are near zero or negative, so the single best cell is plausibly a chance maximum. The paper should report standard errors or bootstrap intervals across repeated training runs and should show whether the best configuration exceeds a null-prediction baseline. As written, the numerical evidence does not support the claim that a specific configuration is 'the right configuration.'
[Section 5.6, Table 6] The claim that PwDRecNet 'outperforms all the regression by large margins' is not meaningful when the absolute correlations are 0.0242 versus about 0.001-0.0019, all essentially zero. The text states that the regression loss 'did not optimize,' which suggests the regression baselines may not have been trained adequately, and no statistical significance is reported. This comparison does not provide evidence that a better network would yield feasible reconstruction.

minor comments (5)

[Table 1] The row for input signal length 1 s appears malformed ('1 + 0.0211 - + 6'); the intended values and any missing entry should be corrected.
[Section 5.4] The text says the results are tabulated in Table 4 but then refers to 'Table 5' in the following sentence; the cross-reference should be corrected.
[Tables 1-6] The use of '+' and '-' to represent values 'close to zero' obscures the actual numbers and prevents reproducibility; exact values should be reported, at least in a supplementary table.
[Section 4] The paper lists MSE as an evaluation metric but no MSE values are reported in any table; either include them or remove the metric.
[Section 3.1] The reference for the three bipolar channel setup is given as [1], but [1] is an iPPG paper; the setup appears to come from the NInFEA-DB paper [12] or the associated PhysioNet record [2], so the citation should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is a standard train/test evaluation against an external public dataset, with no fitted parameter or self-citation used as evidence.

full rationale

The paper's claim is an empirical feasibility statement supported by train/test evaluation on the public NInFEA-DB dataset. The PwD envelopes are extracted from that dataset using a published pipeline [12], the fECG inputs are derived from the same database via a PCA-ICA-PCA pipeline, and PwDRecNet is evaluated on held-out portions under both time-based and random splits. No parameter is fitted to the test set and then relabeled as a prediction; the reported correlation values are actual model outputs on held-out data. The acknowledged lack of explicit time alignment between fECG and PwD (Section 5.1, Figure 4) and the near-flat predicted envelopes are validation and data-quality concerns, not circularity. Likewise, selecting the best configuration from the ablation tables after inspecting results is a statistical selection effect, not a logical reduction of the conclusion to the input. The paper contains no load-bearing self-citation chain: the cited works are external dataset, preprocessing, and prior PwD-classification papers, not the authors' own prior results invoked to force the conclusion. Because the central claim is not equivalent to any input by construction, the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central feasibility claim rests on physiological correspondence between ECG and Doppler flow, on clean fECG extraction, and on time-aligned windowing. These are domain assumptions, not proven here, and the paper itself notes alignment is missing. The model hyperparameters and the choice of best configuration are free parameters selected after seeing test results.

free parameters (5)

Input signal length = 2 s
Chosen as optimum in Table 1 ablation; affects all downstream results.
Batch size = 128
Reported as empirically optimal; used in the best results.
Learning rate and optimizer = 1e-3, RMSprop
Selected empirically; Adam is reported to cause explosive gradients.
Training epochs = 50
Training duration with best model saved; no early stopping analysis.
Best wave and envelope configuration = EA+ / upper envelope / PCA single channel
The headline r=0.0453 is obtained by selecting the best among many configurations post hoc.

assumptions (5)

domain assumption ECG waveform morphology corresponds to ventricular volume and E/A/V Doppler waves (adult schematic adapted from kenhub.com).
Invoked in Section 1 and Figure 1(c) as the physiological motivation; no fetal validation is provided.
domain assumption Fixed-length segmentation of independently resampled fECG and PwD envelopes creates aligned input-target pairs.
Section 3.1 splits signals in fixed time intervals and treats them as respectively coherent; Section 5.1 admits explicit time alignment is absent.
domain assumption PCA-ICA-PCA on three abdominal channels recovers a usable fECG signal.
Section 3.1 relies on this extraction chain; no quantitative fECG quality metric is reported.
domain assumption Envelope extraction from PwD images by Otsu binarization and max-min tracing is a faithful ground truth.
Section 3.1 delegates the method to reference [12]; errors in envelope extraction would propagate into all targets.
domain assumption NInFEA-DB healthy-fetus recordings are representative enough for a general feasibility statement.
Only healthy fetuses are included; generalization to pathology is assumed in the framing.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards reconstruction of Pulsed-wave Doppler signals from Non-invasive fetal ECG." pith.science (2026). https://pith.science/paper/MEK4AC7M

@misc{pith2026250116651,
  author       = {Pith},
  title        = {Pith review of: Towards reconstruction of Pulsed-wave Doppler signals from Non-invasive fetal ECG},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MEK4AC7M}},
  note         = {Machine review of arXiv:2501.16651}
}

read the original abstract

Fetal cardiac health monitoring with invasive methods have a limited viability because they can only be utilized during labor and are uncomfortable. On the other hand non-invasive fECG are adulterated with maternal ECG, and hence resulting in poor analysis. In contrast, Pulsed-wave Doppler (PwD) echocardiography generates high-quality signals representing fetal blood volume inflow-outflow. It also follows non-invasive signal acquisition. The only drawback is that it requires highly expensive setup. To address this aspect, we put forward a challenging research question - can we reconstruct PwD signals using non-invasive fetal ECG?

Figures

Figures reproduced from arXiv: 2501.16651 by the authors.

**Figure 1.** The figure illustrates NI-fECG and PwD signals. It also shows physiological relationship between ECG and PwD for an adult ((c) part of the figure is adapted from https://kenhub.com). prevalent scale. Non-invasive fECG (NI-fECG) has emerged as an alternative for invasive methods. It is collected via multi-electrode set-up placed over maternal abdomen. Although non-invasive, the NI-fECG signal comprises maternal ECG (… view at source ↗

**Figure 2.** The proposed signal preprocessing pipeline. After preprocessing, PwD signal envelopes are used as ground truths, while NI-fECG signals are used inputs for reconstruction. a Bessel bandpass Filter (0.1 – 50 Hz), denoising and removal of artefacts is performed. Filtering the signal in this range further permits to remove primary component (present at 60 Hz frequency). NI-fECG signals are not directly available, they … view at source ↗

**Figure 3.** The proposed PwDRecNet. It follows a UNet style residually connected 1D CNN architecture. It can capture temporal contexts at multiple levels. As illustrated, it’s operation involves taking NI-fECG signal as input and then generating PwD signal envelopes at the output. connected 1D- Convolutions. These three convolutions have residual connection facilitated between them, which is instrumental in preserving temporal … view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Some reconstruction results. ’Blue’ color represents the fECG singal. ’Yellow’ and ’green’ represents ground truth PwD signal’s upper and lower envelope. While, ’maroon’ and ’violet’ depicts the predicted PwD signal’s upper and lower envelope. It is although visually e…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Cross-Modal Generative Framework for Signal Translation from Fetal-Maternal Electrocardiograms to Fetal Doppler Waveforms
cs.LG 2026-07 conditional novelty 6.0 of 10

A cross-modal attention network reconstructs fetal Doppler envelopes from fetal-maternal ECG, showing selective maternal ECG fusion improves frequency-domain fidelity by 39% over naive concatenation.

Reference graph

Works this paper leans on

16 extracted references · 12 canonical work pages · cited by 1 Pith paper

[1]

In: Proc

Comas, A., et al.: Turnip: Time-series u-net with recurrence for nir imaging ppg. In: Proc. IEEE International Conference on Image Processing (ICIP). pp. 309–313 (2021)

work page 2021
[2]

circulation101(23), e215–e220 (2000)

Goldberger, A.L., et al.: Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals. circulation101(23), e215–e220 (2000)

2000
[3]

In: Proc

Luo, J., et al.: Dynamic blood oxygen saturation monitoring based on a new ippg detecting device. In: Proc. International Conference on Biomedical Engineering and Technology. pp. 92–99 (2021)

work page 2021
[4]

Mohebbian, M.R., et al.: Fetal ecg extraction from maternal ecg using attention- basedcyclegan.IEEEJournalofBiomedicalandHealthInformatics 26(2),515–526 (2021)

2021
[5]

Computer Methods and Programs in Biomedicine Update 3, 100096 (2023)

Mohebbian, M.R., et al.: Semi-supervised active transfer learning for fetal ecg arrhythmia detection. Computer Methods and Programs in Biomedicine Update 3, 100096 (2023)

2023
[6]

In: Proc

Nakatani, S., Yamamoto, K., Ohtsuki, T.: Fetal arrhythmia detection based on deep learning using fetal ecg signals. In: Proc. Global Communications Conference (GLOBECOM) (2022)

work page 2022
[7]

Frontiers in Biomedical Technologies10(4), 417–426 (2023)

Rai, R.K., et al.: Fetal ecg arrhythmia detection based on densnet transfer learning. Frontiers in Biomedical Technologies10(4), 417–426 (2023)

2023
[8]

In: Proc

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Proc. Medical Image Computing and Computer- Assisted Intervention (MICCAI). pp. 234–241 (2015)

work page 2015

Show all 16 references

[9]

In: Proc

Sulas, E., et al.: Automatic recognition of complete atrioventricular activity in fetal pulsed-wave doppler signals. In: Proc. Engineering in Medicine and Biology Society (EMBC). pp. 917–920 (2018)

2018
[10]

In: Prof

Sulas, E., et al.: Fetal pulsed-wave doppler atrioventricular activity detection by envelope extraction and processing. In: Prof. Computing in Cardiology (CinC). vol. 45, pp. 1–4 (2018)

2018
[11]

Computer Methods and Programs in Biomedicine 190, 105336 (2020)

Sulas, E., et al.: Automatic detection of complete and measurable cardiac cycles in antenatal pulsed-wave doppler signals. Computer Methods and Programs in Biomedicine 190, 105336 (2020)

2020
[12]

Scientific Data8(1), 30 (2021) Title Suppressed Due to Excessive Length 13

Sulas,E.,etal.:Anon-invasivemultimodalfoetalecg–dopplerdatasetforantenatal cardiology research. Scientific Data8(1), 30 (2021) Title Suppressed Due to Excessive Length 13

2021
[13]

IEEE Transactions on Instrumentation and Measurement (2023)

Wang, X., He, Z., Lin, Z., Han, Y., Su, W., Xie, S.: Correlation-aware attention cyclegan for accurate fetal ecg extraction. IEEE Transactions on Instrumentation and Measurement (2023)

2023
[14]

In: Proc

Zhu, Q., et al.: Ecg reconstruction via ppg: A pilot study. In: Proc. International Conference on Biomedical & Health Informatics (BHI). pp. 1–4 (2019)

2019
[15]

IEEE Internet of Things Journal8(23), 16734–16748 (2021)

Zhu, Q., et al.: Learning your heart actions from pulse: Ecg waveform reconstruc- tion from ppg. IEEE Internet of Things Journal8(23), 16734–16748 (2021)

2021
[16]

IEEE Transactions on Biomedical Engineering69(2), 830–839 (2021)

Zschocke, J., et al.: Reconstruction of pulse wave and respiration from wrist ac- celerometer during sleep. IEEE Transactions on Biomedical Engineering69(2), 830–839 (2021)

2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

In: Proc

Comas, A., et al.: Turnip: Time-series u-net with recurrence for nir imaging ppg. In: Proc. IEEE International Conference on Image Processing (ICIP). pp. 309–313 (2021)

work page 2021

[2] [2]

circulation101(23), e215–e220 (2000)

Goldberger, A.L., et al.: Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals. circulation101(23), e215–e220 (2000)

2000

[3] [3]

In: Proc

Luo, J., et al.: Dynamic blood oxygen saturation monitoring based on a new ippg detecting device. In: Proc. International Conference on Biomedical Engineering and Technology. pp. 92–99 (2021)

work page 2021

[4] [4]

Mohebbian, M.R., et al.: Fetal ecg extraction from maternal ecg using attention- basedcyclegan.IEEEJournalofBiomedicalandHealthInformatics 26(2),515–526 (2021)

2021

[5] [5]

Computer Methods and Programs in Biomedicine Update 3, 100096 (2023)

Mohebbian, M.R., et al.: Semi-supervised active transfer learning for fetal ecg arrhythmia detection. Computer Methods and Programs in Biomedicine Update 3, 100096 (2023)

2023

[6] [6]

In: Proc

Nakatani, S., Yamamoto, K., Ohtsuki, T.: Fetal arrhythmia detection based on deep learning using fetal ecg signals. In: Proc. Global Communications Conference (GLOBECOM) (2022)

work page 2022

[7] [7]

Frontiers in Biomedical Technologies10(4), 417–426 (2023)

Rai, R.K., et al.: Fetal ecg arrhythmia detection based on densnet transfer learning. Frontiers in Biomedical Technologies10(4), 417–426 (2023)

2023

[8] [8]

In: Proc

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Proc. Medical Image Computing and Computer- Assisted Intervention (MICCAI). pp. 234–241 (2015)

work page 2015